Archives Discussions

[spectral]

Hi,

I would like to know what is the difference between :

oclBuff = clCreateBuffer(gpu_context, CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR, mem_size, dataSet, &err);

...and...

oclBuff = clCreateBuffer(gpu_context, CL_MEM_READ_ONLY, mem_size, NULL, &err);

err = clEnqueueWriteBuffer(cmd_queue, dataSet, CL_TRUE, 0, mem_size, a, 0, NULL, NULL);

Which one I should use in which case ? and why ? Or it is exactly the same ?

Thanks

[genaganna]

viewon01,

       In both cases, you are creating buffer and initializing the buffer with your own data.  only thing is that you are using two functions to do that in the second case.

       One difference i can tell you.   Choose first case if you want create buffer and initialize only once in whole program.   Use clEnqueueWriteBuffer if you want to update the buffer values with new data set more than once.

[spectral]

Thanks for the precision,

Regards

[n_treutner]

Originally posted by: genaganna viewon01,

       In both cases, you are creating buffer and initializing the buffer with your own data.  only thing is that you are using two functions to do that in the second case.

       One difference i can tell you.   Choose first case if you want create buffer and initialize only once in whole program.   Use clEnqueueWriteBuffer if you want to update the buffer values with new data set more than once.

Your hint, what the difference is, just fixed a memory leak, that i've been working on for 2 weeks now.

Thanks!

[davibu]

Originally posted by: viewon01

oclBuff = clCreateBuffer(gpu_context, CL_MEM_READ_ONLY, mem_size, NULL, &err);

err = clEnqueueWriteBuffer(cmd_queue, dataSet, CL_TRUE, 0, mem_size, a, 0, NULL, NULL);

Does clCreateBuffer() include an implicit call to clEnqueueWriteBuffer() too ?

I mean, for a read only buffer, is it required to call clEnqueueWriteBuffer() after the creation of the buffer of it is implicitaly done buy clCreateBuffer() ?

Thanks,

David

[genaganna]

Originally posted by: davibu

Originally posted by: viewon01

oclBuff = clCreateBuffer(gpu_context, CL_MEM_READ_ONLY, mem_size, NULL, &err);

err = clEnqueueWriteBuffer(cmd_queue, dataSet, CL_TRUE, 0, mem_size, a, 0, NULL, NULL);

Does clCreateBuffer() include an implicit call to clEnqueueWriteBuffer() too ?

        No

I mean, for a read only buffer, is it required to call clEnqueueWriteBuffer() after the creation of the buffer of it is implicitaly done buy clCreateBuffer() ?

You need to call clEnqueueWriteBuffer if CL_MEM_COPY_HOST_PTR is not used in clCreateBuffer.  Please read spec for more details.

[davibu]

Originally posted by: genaganna

You need to call clEnqueueWriteBuffer if CL_MEM_COPY_HOST_PTR is not used in clCreateBuffer.  Please read spec for more details.

Thank you very much.

[_Big_Mac_]

What if there are multiple devices within the context? clCreateBuffer only takes a context as a parameter, not a command queue associated with a specific device. I've been told on Nvidia forums that clCreateBuffer with CL_MEM_COPY_HOST_PTR will copy the data to ALL devices associated with that context. So if you're only using a single device, you might be better off manually enqueuing a write to this particular device because if a user's machine has several GPUs, he will pay for several memcopies over PCIE. Is this correct?

[tal1974]

Thank you. It's very important addition.